It is possible for an async thread to simply be deferred if not
specifically set to async. In which case, until the return value is
requested, the loop may simply wait.
By decoupling from std::thread, we can avoid the wait/sleep cycle in
checking the std::atomic completion variable. The std::future variables
are immediately returning without the additional atomic check cycle.
This minimizes the thread usage, particularly during online actions such
as moving items and routing.
UpdateRatsnest: Minimized the number of atomics by utilizing future
promise. Avoided updating node-less nets. Node-less nets are common in
the dynamic ratsnest because the net vector is created from the board
nets by value but only populated with a subset. So the dynamic ratsnest
calculation always saw every net as dirty and spun up extra threads.
searchConnections: ParallelThreadCount is no longer lower bounded.
Small item counts are handled inline rather than with a separate thread.
Some windows machines seem prone to sleeping for excessive times. When
called in the main thread, this may prevent the system from responding
to Windows 'Are you alive?' calls.
Rather than forcing each thread to iterate over the full nets list to
find the dirty items, we collect the dirty items first before asking the
threads to address them. This prevents thread spinning. We also
amortize the cost of calculating on an 8-net per thread basis rather
than 2.
Locks protect the std::set in each item. Devolving the mutex to the
CN_ITEM allows multiple threads to make simultaneous connections to
different items where they do not conflict.
The connectivity files were unwieldy. This separates them logically
into data, algo and items where the items classes are those that hold,
surprise, surprise, the items, lists and clusters.
Seth Hillbrand
jean-pierre charras